Talking about ways to use an off-the-shelf language model to solve a multiple-choice task. Covering:
- Intro to the Kaggle competition
- Benchmarking with GPT3.5
- Using the OpenAI function calling API to enforce structure on answers
- Using Llama2 as a classifier by examining the logits (next token predictions)
- Using perplexity to evaluate question-answer pairs
Notebook using the OpenAI API to test GPT3.5: www.kaggle.com/johnowhitaker/...
Llama2 demo notebook: colab.research.google.com/dri... (quickly made for this video, don't trust the calculations, rather start with the below notebook)
Notebook testing different open models with the perplexity approach: www.kaggle.com/code/takamichi... (a good template to start experimenting since it shows how to run as a submission.
6 авг 2023